Ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU

Abstract

The invention belongs to the field of wind power prediction, and provides an ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU by considering the strong volatility and randomness of wind power. The method comprises the following steps: firstly, constructing a secondary mode decomposition framework based on improved complete set empirical mode decomposition, sample entropy, K-means clustering and variational mode decomposition, adaptively selecting components with relatively high complexity for secondary decomposition, and weakening volatility and randomness of an original sequence; and then, on the basis of the time convolutional network, an efficient channel attention module (ECA) is introduced to mine a coupling relationship among components, weights are given to the components in a self-adaptive manner to highlight important information, and a long-term time sequence dependency relationship of the components is further captured by using a bidirectional gating cycle unit, so that the prediction effect of the ultra-short-term wind power is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of short-term wind power prediction, and more specifically relates to an ultra-short-term wind power prediction method based on quadratic decomposition and TCN-ECA-BiGRU, and in particular to the prediction of ultra-short-term wind power that is strongly correlated with time. Background Art

[0002] In order to cope with the increasingly severe energy shortage and environmental degradation, global energy is transforming from fossil fuels to clean energy such as wind power. Although wind power has great potential, the intermittent and volatile nature of wind resources poses great challenges to the planning, operation and control of wind farms. In addition, the large-scale grid connection of wind power also endangers the balanced operation of the power system. Therefore, accurate prediction of wind power is crucial for the reasonable arrangement of system backup, ensuring the stable operation of the power grid and improving economic benefits.

[0003] At present, the prediction methods of wind power are mainly divided into physical methods, statistical methods and machine learning methods. The physical method uses factors such as terrain and weather to establish a physical prediction model, and combines the wind speed and power characteristic curve of the wind turbine to predict the output power of the wind farm. However, the modeling process of the physical method is complex, the calculation speed is slow, and the accuracy is difficult to guarantee, which limits its development. Statistical methods include multivariate linear regression and Kalman filtering. These methods have simple models, small calculations, and fast prediction speeds, but poor robustness and are easily affected by random factors, resulting in low reliability in complex wind power prediction. Machine learning methods such as extreme learning machines (ELM) and backpropagation neural networks (BPNN) can handle nonlinear and multidimensional data well, but are sensitive to outliers and are prone to model performance degradation when faced with data mutations or breakpoints.

[0004] As a branch of machine learning, deep learning is widely regarded as the most promising solution for extracting complex nonlinear relationships from massive data. Compared with traditional methods such as BPNN and ELM, deep learning methods have a deeper network structure, can capture potential relevant information in high-dimensional state space, and show superior prediction performance. Some scholars have established a long short-term memory (LSTM) network based on the attention mechanism to extract time series features from historical data. Based on this, a hybrid model combining convolutional neural network (CNN) and LSTM network is proposed. CNN is used to initially extract multidimensional features, and then these features are input into the LSTM network, which effectively improves the prediction accuracy. However, due to the high volatility of wind power series, it is still difficult to use a single method to predict wind power. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention proposes a short-term wind power prediction method based on quadratic decomposition and TCN-ECA-BiGRU, which is used to solve the problem that the existing model cannot capture the coupling relationship between the characteristics of each component and the prediction effect is poor.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A short-term wind power forecasting method based on quadratic decomposition and TCN-ECA-BiGRU, the method comprising the following steps:

[0008] Step 1: Analyze the temporal and spatial characteristics of wind power data and preprocess the data;

[0009] Step 2: Use ICEEMDAN to decompose the original wind power series and calculate the sample entropy values ​​of all IMF components after the decomposition.

[0010] Step 3: Use the K-means clustering algorithm based on sample entropy to divide the IMF components decomposed by ICEEMDAN into two categories and reconstruct them into high-frequency components and low-frequency components. Then, the high-frequency components are further decomposed using VMD.

[0011] Step 4: Construct the TCN-ECA-BiGRU prediction model and input all components obtained through the secondary decomposition into the prediction model to achieve ultra-short-term prediction of wind power.

[0012] Step 5: Perform error assessment on the ultra-short-term wind power forecast results in step 4.

[0013] The technical solution is further optimized, and the wind power data preprocessing in step 1 is as follows:

[0014] The historical wind power data of wind farms at different times are collected according to a fixed period within a certain time range, and the missing values ​​in the original wind power data are filled by linear interpolation method, and the wind power data are normalized by Min-Max.

[0015] The step 2 is as follows:

[0016] Step 2.1: Use ICEEMDAN to decompose the wind power data:

[0017] Wind power data is affected by factors such as weather and season, which makes the data nonlinear and volatile, and direct prediction has certain limitations. Improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is a time-frequency analysis method with adaptive characteristics. The algorithm decomposes the wind power sequence into a set of intrinsic mode functions (IMFs) with frequencies from high to low, and each IMF reflects the characteristics of the original wind power data in different frequency ranges.

[0018] ICEEMDAN is an improvement on CEEMDAN, solving the pseudo-modal problem generated in its decomposition process. Compared with traditional empirical mode decomposition and ensemble empirical mode decomposition, the improved algorithm more effectively solves the modal aliasing problem by adding non-Gaussian white noise, further reducing the impact of residual noise.

[0019] Step 2.2: Calculate the sample entropy of all IMF components:

[0020] Sample Entropy (SE) is an algorithm for measuring the complexity of time series. Compared with approximate entropy, sample entropy has better consistency and its calculation does not depend on the length of data. Sample entropy is used to measure the complexity of each subsequence obtained by ICEEMDAN decomposition. Generally speaking, the larger the value of sample entropy, the higher the complexity of the sequence, and vice versa.

[0021] The step 3 is as follows:

[0022] Step 3.1: Use the K-means algorithm to cluster the IMF components based on the sample entropy values ​​obtained in step 2 into two categories: high-frequency components and low-frequency components, and reconstruct these two types of components.

[0023] Step 3.2: Perform mode decomposition again on the high-frequency components in step 3.1 using variational mode decomposition (VMD), combine the IMF components obtained by the secondary decomposition with the previous low-frequency components to obtain the final intrinsic mode function component set.

[0024] The step 4 is as follows:

[0025] Step 4.1: Improve the efficient channel attention module (ECA)

[0026] The 2D global average pooling layer in the ECA module is replaced with a 1D global average pooling layer to compress time series data of different dimensions. The improved EC can operate on a large number of features extracted by the TCN module and assign weights to different features. Finally, the original input features are combined with the channel weights to obtain a weighted sequence of feature maps. Important channel features are given larger weights to achieve enhancement, and invalid channel features are given smaller weights to achieve suppression.

[0027] Step 4.2: Build TCN-ECA-BiGRU prediction model:

[0028] First, the ECA module is integrated into the TCN residual block. The TCN residual block extracts the time features of the multivariate time series. The ECA module performs an average global pooling (GAP) operation on the extracted features to obtain all features without dimensionality reduction, and emphasizes the importance of different features through different weight assignments. Then, BiGRU is used to further capture and learn long-term temporal dependencies, improve the generalization ability of the model and avoid the limitations of a single model. Finally, the mapping relationship between features and outputs is established through the fully connected layer to achieve high-precision prediction of wind power.

[0029] The step 5 is as follows:

[0030] Step 5: Based on the prediction results in step 4.2, perform error evaluation according to the error calculation formula.

[0031] In view of the strong volatility of wind power, the difficulty in adaptively selecting the IMF components that need to be further decomposed after one decomposition, and the inability to effectively capture the coupling relationship between the IMF components, the present invention proposes a TCN-ECA-BiGRU model based on adaptive quadratic mode decomposition. This method uses an adaptive quadratic mode decomposition method to decompose the wind power sequence, which can effectively solve the problem that conventional quadratic mode decomposition is difficult to determine which IMF components need to be reconstructed, weaken the complexity of high-frequency components, and obtain smoother and more stable components. Combining TCN, ECA and BiGRU networks can effectively adapt to input features, realize local cross-channel interaction, capture the coupling relationship between the IMF components, enhance the ability to capture important information, and stably capture the long-term dependencies of time series, thereby effectively improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 To improve the basic structure diagram of the ECA module;

[0033] Figure 2 This is the basic structure diagram of the TCN-ECA-BiGRU prediction model;

[0034] Figure 3 This is the prediction model diagram. DETAILED DESCRIPTION

[0035] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments.

[0036] The present invention takes into account that wind power has strong volatility and randomness, and the signal decomposition method can reduce data complexity and improve data stability. The dilated causal convolution structure of the TCN network can increase the receptive field, which helps the model learn the medium- and long-term dependencies of the time series. The ECA module can operate on a large number of features extracted by the TCN module and assign weights to different features. BiGRU can usually better grasp the global time series information, so that the model can obtain better prediction performance. Therefore, the present invention proposes a short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU.

[0037] The method specifically comprises the following steps:

[0038] Step 1: Analyze the temporal and spatial characteristics of wind power data and preprocess the data. Collect historical wind power data of wind farms at different times in a fixed period within a certain time range, use linear interpolation to fill in the missing values ​​in the original wind power data, and perform Min-Max normalization on the wind power data.

[0039] The photovoltaic data preprocessing steps are as follows:

[0040] Step 1.1: Wind power data collection:

[0041] The historical wind power data of the wind farm at different times is collected according to a fixed period within a certain time range.

[0042] Step 1.2: Wind power data preprocessing:

[0043] Linear interpolation is used to fill the missing values ​​in the original wind power data, that is, the average value of the data at the previous moment and the next moment is used to replace the missing data at this moment.

[0044] In order to improve the convergence speed and stability of the model, the following formula is used to perform Min-Max normalization on the data.

[0045]

[0046] Where: X i is the normalized wind power value; X min is the minimum value in the wind power sequence, X max is the maximum value in the wind power series.

[0047] Step 2: Use ICEEMDAN to decompose the original wind power series and calculate the sample entropy values ​​of all IMF components after the decomposition.

[0048] The process of using ICEEMDAN to decompose the original wind power series and calculate the sample entropy of all IMF components after the decomposition is as follows:

[0049] Step 2.1: Use ICEEMDAN to decompose the original wind power series.

[0050] Wind power data is affected by factors such as weather and season, which makes the data nonlinear and volatile, and direct prediction has certain limitations. Improved complete ensemble empirical mode decomposition (ICEEMDAN) is a time-frequency analysis method with adaptive characteristics. The algorithm decomposes the wind power sequence into a set of intrinsic mode functions (IMFs) with frequencies from high to low, and each IMF reflects the characteristics of the original wind power data in different frequency ranges.

[0051] ICEEMDAN is an improvement on CEEMDAN, solving the pseudo-modal problem generated in its decomposition process. Compared with traditional empirical mode decomposition and ensemble empirical mode decomposition, the improved algorithm more effectively solves the modal aliasing problem by adding non-Gaussian white noise, further reducing the impact of residual noise.

[0052] The steps of ICEEMDAN to decompose the original wind power sequence are as follows:

[0053] 1) Let x be the original wind power time series, and the special noise E k (w (i) ) is added to x:

[0054] x (i) =x+β 1 E k (w (i) ) (16)

[0055] Where: x (i) is the wind power time series after adding the i-th group of special noise; β 1 is the expected signal-to-noise ratio of the first decomposition iteration; E k (·) is the k-order modal component generated by empirical mode decomposition; w (i) represents the i-th group of white noise used to assist decomposition.

[0056] 2) Calculate the residual r of the first decomposition 1 , and the modal component d is obtained by subtracting the residual from the original wind power time series x 1 .

[0057]

[0058] d 1 =xr 1 (18)

[0059] Where: M(·) is the operator that generates the local mean of the signal; I is the number of white noise groups.

[0060] 3) Continue to add white noise signal and calculate the second residual r 2 , find the second-order modal component d of the original wind power time series 2 .

[0061]

[0062] d 2 =r 1 -r 2 (20)

[0063] 4) Similar to step 3), calculate the kth residual and modal component.

[0064]

[0065] d k =r k-1 -r k (twenty two)

[0066] 5) Repeat step 4) until the maximum number of iterations is reached or the residual cannot be decomposed any further.

[0067] Step 2.2: Calculate the sample entropy of all IMF components after one decomposition:

[0068] Sample Entropy (SE) is an algorithm for measuring the complexity of time series. Compared with approximate entropy, sample entropy has better consistency and its calculation does not depend on the length of data. Sample entropy is used to measure the complexity of each subsequence obtained by ICEEMDAN decomposition. Generally speaking, the larger the value of sample entropy, the higher the complexity of the sequence, and vice versa.

[0069] According to formula (8), the complexity of each IMF component decomposed by ICEEMDAN is measured

[0070]

[0071] Where: Sampen(·) is the SE calculation function; N is the length of the modal component; B m (r), B m+1 (r) are the probabilities of the sequence matching m points and m+1 points under tolerance r. In practical scenarios, m is the dimension, which is generally 2. r is the tolerance, which is usually 10% to 25% of the standard deviation (SD) of the calculated data. In this paper, r is 20% of SD.

[0072] Step 3: Use the K-means clustering algorithm based on sample entropy to divide the IMF components decomposed by ICEEMDAN into two categories and reconstruct them into high-frequency components and low-frequency components. Then, VMD is used to further decompose the high-frequency components. The specific steps are as follows:

[0073] Step 3.1: Use the K-means algorithm to cluster the IMF components based on the sample entropy values ​​obtained in step 2 into two categories: high-frequency components and low-frequency components, and reconstruct these two types of components.

[0074] The steps of clustering the IMF components decomposed by ICEEMDAN using the K-means clustering algorithm based on sample entropy are as follows:

[0075] 1) Randomly select k samples as the initial cluster centers;

[0076] 2) Measure the distance between the cluster center and the remaining objects by Euclidean distance and assign each object to the nearest center.

[0077] 3) Take the average value of each category as the new cluster center.

[0078] 4) Iterate steps 2) and 3) until the new cluster center no longer changes or the maximum number of iterations is reached.

[0079] In step 2, all IMF components are divided into two categories, one is high frequency components and the other is low frequency components. Then these two categories of components are reconstructed.

[0080] Step 3.2: Use the VMD algorithm to perform modal decomposition again on the high-frequency components in step 3.1, and combine the IMF components obtained by the secondary decomposition with the previous low-frequency components to obtain the final intrinsic mode function component set.

[0081] The steps for further decomposing the high-frequency components using VMD are as follows:

[0082] 1) Construct a constrained variational optimal problem for the intrinsic mode function:

[0083]

[0084] Where: K is the number of decomposed modes; u k 、w k are the kth subcomponent and its center frequency after the high-frequency non-stationary component IMF1 is decomposed; u k (t) is the value of the kth modal component at time t; * is the convolution operator; represents the square of the two-norm; Denotes partial derivative; δ t is the Dirac function; f(t) is the high-frequency strong non-stationary component decomposed from the original wind power sequence.

[0085] 2) Add the penalty term α and the Lagrange multiplication operator λ to transform the constrained variational problem into an unconstrained variational model:

[0086]

[0087] Where: L(·) is the Lagrangian function; <·> represents the inner product.

[0088] 3) Using the alternating direction multiplier iteration algorithm and Fourier isometric transform to calculate u k and w k Optimize and judge the accuracy of each iteration. When the required accuracy is achieved, the iteration outputs each component.

[0089] Step 4: Construct the TCN-ECA-BiGRU prediction model and input all components obtained through the secondary decomposition into the prediction model to achieve ultra-short-term prediction of wind power.

[0090] Step 4.1: Improve the ECA module:

[0091] The two-dimensional global average pooling layer in the ECA module is replaced with a one-dimensional global average pooling layer to compress time series data of different dimensions. The improved ECANet can operate on a large number of features extracted by the TCN module and assign weights to different features. Finally, the original input features are combined with the channel weights to obtain a weighted sequence of feature maps. Important channel features are given greater weights to achieve enhancement, and invalid channel features are given smaller weights to achieve suppression. The ECA module structure here is as follows Figure 1 As shown, where T is the time step of the sequence, C is the different dimensions of the sequence, U is the original sequence, GAP is the global average pooling layer, and k not only represents the size of the one-dimensional convolution kernel, but also the frequency of local cross-channel interactions. is the weighted sequence.

[0092] In order to apply ECANet to deep TCN, the two-dimensional global average pooling layer needs to be replaced with a one-dimensional global average pooling layer to compress time series data of different dimensions. The improved ECANet can operate on a large number of features extracted by the TCN module and assign weights to different features.

[0093] The ECA module first performs an average global pooling (GAP) operation on the original input feature sequence to obtain all features without dimensionality reduction. Then, in order to avoid manually adjusting the k value, a method is introduced to adaptively determine the interaction coverage k according to the size of the input channel C. The calculation formula is

[0094]

[0095] Where: |t odd is the odd number closest to t; C is the current total number of channels; γ and b are fixed values ​​2 and 1 respectively. After the k value is adaptively determined, in order to further simplify the model, all channels share the learned attention weight parameters, then the i-th channel y i The weight w i It can be expressed as:

[0096]

[0097] Where: For i The set of k adjacent channels; α j is the weight parameter shared by all channels; σ is the sigmoid function.

[0098] Finally, the original input features are combined with the channel weights to obtain a weighted sequence of feature maps Important channel features are given larger weights to achieve enhancement, and invalid channel features are given smaller weights to achieve suppression.

[0099] Step 4.2: Build TCN-ECA-BiGRU prediction model:

[0100] TCN-ECA-BiGRU prediction model Figure 2 As shown in the figure. First, the ECA module is integrated into the TCN residual block. The TCN residual block extracts the time features of the multivariate time series. The ECA module performs an average global pooling (GAP) operation on the extracted features to obtain all features without dimensionality reduction, and emphasizes the importance of different features through different weight assignments. Then, BiGRU is used to further capture and learn long-term temporal dependencies, improve the generalization ability of the model and avoid the limitations of a single model. Finally, the mapping relationship between features and outputs is established through the fully connected layer to achieve high-precision prediction of wind power.

[0101] Step 4.3: Construct an ultra-short-term wind power prediction model based on quadratic mode decomposition and TCN-ECA-BiGRU:

[0102] Prediction models such as Figure 3 As shown in the figure. First, the preprocessed wind power sequence is decomposed using ICEEMDAN to obtain multiple subsequence components (IMF). Since the high-frequency components in the IMF decomposed by ICEEMDAN are highly complex, the prediction effect is poor. In order to improve the prediction effect of the model and reduce the calculation amount of the subsequent training model, the sample entropy is used to evaluate the complexity of each IMF component, and the K-means clustering method is used to cluster the IMF components according to the sample entropy values ​​of each IMF component, and reconstruct them into high-frequency components (CIMF0) and low-frequency components (CIMF1). Then, the high-frequency components are further decomposed based on VMD to obtain multi-scale VIMF components reflecting different characteristics such as wind power trend, periodicity and randomness. The decomposed VIMF components and low-frequency components are combined to form the final IMF component set as the model input; finally, the TCN model combined with the ECA module is used to realize multi-scale component feature learning, fully extract the coupling relationship between each IMF component, and combine TCN-ECA with the BiGRU model to further capture and learn long-term time dependencies, improve the generalization ability of the model and avoid the limitations of a single model.

[0103] Step 5: Based on the prediction results in step 4.2, perform error evaluation according to the error calculation formula.

[0104] Step 5: The error assessment process of the ultra-short-term wind power forecast results is as follows:

[0105] Calculate the mean absolute error index E MAE , root mean square error index E RMSE and the mean absolute percentage error E MAPEThe error of the prediction result is evaluated.

[0106] The formulas are:

[0107]

[0108] Where: Y i and Y i are the true value and predicted value of the i-th test sample respectively; N is the total number of test samples. MAE It represents the mean absolute value of the error between the predicted value and the true value, E RMSE Indicates the absolute value of the error between the predicted value and the true value, E MAPE Represents the absolute value of the error between the predicted value and the true value. MAE 、E RMSE and E MAPE The smaller the value, the higher the prediction accuracy.

[0109] Compared with the prior art, the beneficial effects of the present invention are mainly manifested in:

[0110] In view of the volatility and randomness of wind power, an ultra-short-term wind power forecasting method based on secondary decomposition and TCN-ECA-BiGRU is proposed. In view of the problem of high-frequency modal components in the primary decomposition, a secondary decomposition framework based on ICEEMDAN, K-means and VMD is designed to better reduce data complexity and improve data stability. In view of the problem that traditional forecasting methods cannot effectively capture the coupling relationship between various IMF components, a forecasting model based on TCN, ECA and BIGRU is designed, which can effectively adapt to input features, realize local cross-channel interaction, capture the coupling relationship between various IMF components, enhance the ability to capture important information, and stably capture the long-term dependencies of time series.

[0111] It should be noted that, in this article, terms such as "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "include..." or "include..." do not exclude the existence of additional elements in the process, method, article or terminal device including the elements. In addition, in this article, "greater than", "less than", "exceed" and the like are understood to exclude the number itself; "above", "below", "within" and the like are understood to include the number itself.

[0112] Although the above embodiments have been described, once those skilled in the art know the basic creative concepts, they can make additional changes and modifications to these embodiments. Therefore, the above description is only an embodiment of the present invention and does not limit the patent protection scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the specification and drawings of the present invention, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An ultra-short-term wind power forecasting method based on quadratic decomposition and TCN-ECA-BiGRU is characterized by: The following steps are involved: Step 1: Analyze the temporal and spatial characteristics of wind power data and preprocess the data; Step 2: Use ICEEMDAN to decompose the original wind power series and calculate the sample entropy values ​​of all IMF components after the decomposition. Step 3: Use the K-means clustering algorithm based on sample entropy to divide the IMF components decomposed by ICEEMDAN into two categories and reconstruct them into high-frequency components and low-frequency components. Then, the high-frequency components are further decomposed using VMD. Step 4: Build a TCN-ECA-BiGRU prediction model, input all components obtained through secondary decomposition into the prediction model, and realize ultra-short-term prediction of wind power; Step 5: Perform error assessment on the ultra-short-term wind power forecast results in step 4.

2. The ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU according to claim 1 is characterized in that: The step 1 specifically includes: Step 1.1: Wind power data collection: Collect historical wind power data of wind farms at different times within a certain time range according to fixed cycles; Step 1.2: Wind power data preprocessing: Linear interpolation is used to fill in the missing values ​​in the original wind power data, that is, the average value of the data at the previous moment and the next moment is used to replace the missing data at this moment. In order to improve the convergence speed and stability of the model, the following formula is used to perform Min-Max normalization on the data: Where: X i is the normalized wind power value; X min is the minimum value in the wind power sequence, X max is the maximum value in the wind power series.

3. The ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU according to claim 1 is characterized in that: The step 2 specifically includes: Step 2.1: Use ICEEMDAN to decompose the original wind power sequence, including the following steps: 1) Let x be the original wind power time series, and the special noise E k (w (i) ) is added to x: x (i) =x+β1E k (w (i) ) (1) Where: x (i) is the wind power time series after adding the i-th group of special noise; β1 is the iterative expected signal-to-noise ratio of the first decomposition; E k (·) is the k-order modal component generated by empirical mode decomposition; w (i) represents the i-th group of white noise used to assist decomposition; 2) Calculate the residual r1 of the first decomposition and subtract the residual from the original wind power time series x to obtain the modal component d1. d1=x-r1 (3) Where: M(·) is the operator that generates the local mean of the signal; I is the number of white noise groups; 3) Continue to add white noise signal, calculate the second residual r2, and find the second-order modal component d2 of the original wind power time series. d2＝r1-r2 (5) 4) As in step 3), calculate the kth residual and modal component, d k ＝r k-1 -r k (7) 5) Repeat step 4) until the maximum number of iterations is reached or the residual cannot be decomposed any further; Step 2.2: Calculate the sample entropy of all IMF components after one decomposition. According to formula (8), the complexity of each IMF component decomposed by ICEEMDAN is measured Where: Sampen(·) is the SE calculation function; N is the length of the modal component; B m (r), B m+1 (r) are the probabilities of the sequence matching m points and m+1 points under tolerance r, respectively.

4. The ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU according to claim 1 is characterized in that: The step 3 specifically includes: Step 3.1: Use the K-means clustering algorithm based on sample entropy to cluster the IMF components decomposed by ICEEMDAN. The specific steps are as follows: 1) Randomly select k samples as the initial cluster centers; 2) Measure the distance between the cluster center and the remaining objects by Euclidean distance and assign each object to the nearest center; 3) Take the average value of each category as the new cluster center; 4) Iterate steps 2) and 3) until the new cluster center no longer changes or the maximum number of iterations is reached; In step 2, all IMF components are divided into two categories, one is high-frequency components and the other is low-frequency components, and then these two categories of components are reconstructed; Step 3.2: VMD is used to further decompose the high-frequency components. The specific steps are as follows: 1) Construct a constrained variational optimal problem for the intrinsic mode function: Where: K is the number of decomposed modes; u k 、w k are the kth subcomponent and its center frequency after the high-frequency non-stationary component IMF1 is decomposed; u k (t) is the value of the kth modal component at time t; * is the convolution operator; represents the square of the two-norm; Denotes partial derivative; δ t is the Dirac function; f(t) is the high-frequency strong non-stationary component decomposed from the original wind power sequence; 2) Add the penalty term α and the Lagrange multiplication operator λ to transform the constrained variational problem into an unconstrained variational model: Where: L(·) is the Lagrangian function; <·> represents the inner product; 3) Using the alternating direction multiplier iteration algorithm and Fourier isometric transform to calculate u k and w k Optimize and judge the accuracy of each iteration. When the required accuracy is achieved, the iteration outputs each component.

5. The ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU according to claim 1, characterized in that: The step 4 specifically includes: Step 4.1: Improve the ECA module: The ECA module first performs an average global pooling operation on the original input feature sequence to obtain all features without dimensionality reduction. Then, in order to avoid manually adjusting the k value, a method is introduced to adaptively determine the interaction coverage k according to the size of the input channel C. The calculation formula is: Where: |t odd is the odd number closest to t; C is the current total number of channels; γ and b are fixed values ​​2 and 1 respectively. After the k value is adaptively determined, in order to further simplify the model, all channels share the learned attention weight parameters, then the i-th channel y i The weight w i It can be expressed as: Where: For i The set of k adjacent channels; α j is the weight parameter shared by all channels; σ is the sigmoid function, Finally, the original input features are combined with the channel weights to obtain a weighted sequence of feature maps Assign larger weights to important channel features to achieve enhancement, and smaller weights to invalid channel features to achieve suppression; Step 4.2: Build TCN-ECA-BiGRU prediction model: Firstly, the TCN residual block integrated with the ECA module is used to extract the spatiotemporal features of the multivariate time series, and the importance of different features is emphasized by different weight assignments. Then, BiGRU is used to further capture and learn long-term temporal dependencies. Finally, the connection between features and outputs is established through the fully connected layer to achieve wind power prediction. Step 4.3: Construct an ultra-short-term wind power prediction model based on quadratic mode decomposition and TCN-ECA-BiGRU: Firstly, the preprocessed wind power sequence is decomposed by ICEEMDAN to obtain multiple subsequence components (IMF). Since the high-frequency components in the IMF decomposed by ICEEMDAN are of high complexity, the prediction effect is poor. In order to improve the prediction effect of the model and reduce the calculation amount of the subsequent training model, the complexity of each IMF component is evaluated by sample entropy, and the IMF components are clustered according to the sample entropy values ​​of each IMF component using the K-means clustering method, and reconstructed into high-frequency components (CIMF0) and low-frequency components (CIMF1). Then, the high-frequency components are further decomposed based on VMD to obtain multi-scale VIMF components reflecting different characteristics such as trend, periodicity and randomness of wind power. The decomposed VIMF components and low-frequency components are combined to form the final IMF component set as the model input; finally, the TCN model combined with the ECA module is used to realize multi-scale component feature learning, fully extract the coupling relationship between each IMF component, and combine TCN-ECA with the BiGRU model to further capture and learn long-term time dependencies, improve the generalization ability of the model and avoid the limitations of a single model.

6. The ultra-short-term wind power prediction method based on secondary decomposition and TCN-ECA-BiGRU according to claim 1, characterized in that: The step 5 specifically includes: Calculate the mean absolute error index E MAE , root mean square error index E RMSE and the mean absolute percentage error E MAPE The value of is used to evaluate the error of the prediction results. The formulas are: Where: Y i and Y i are the true value and predicted value of the i-th test sample respectively; N is the total number of test samples, where E MAE It represents the mean absolute value of the error between the predicted value and the true value, E RMSE Represents the absolute value of the error between the predicted value and the true value, E MAPE Represents the absolute value of the error between the predicted value and the true value. MAE 、E RMSE and E MAPE The smaller the value, the higher the prediction accuracy.

Images